717   v16u8 src0, src1, src2, src3, src4, src5, src6, src7;  in SSE16x16_MSA()  local
721   LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);  in SSE16x16_MSA()
725   PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3);  in SSE16x16_MSA()
729   LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);  in SSE16x16_MSA()
733   PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3);  in SSE16x16_MSA()
744   v16u8 src0, src1, src2, src3, src4, src5, src6, src7;  in SSE16x8_MSA()  local
748   LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);  in SSE16x8_MSA()
752   PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3);  in SSE16x8_MSA()
763   v16u8 src0, src1, src2, src3, src4, src5, src6, src7;  in SSE8x8_MSA()  local
768   LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7);  in SSE8x8_MSA()
772   ILVR_B4_UB(src4, src5, src6, src7, ref4, ref5, ref6, ref7, t0, t1, t2, t3);  in SSE8x8_MSA()